Datasets:

Modalities:
Text
Languages:
English
ArXiv:
Libraries:
Datasets
License:
premise
stringlengths
5
1.82k
hypothesis
stringlengths
1
350
label
class label
3 classes
idx
int32
0
393k
Conceptually cream skimming has two basic dimensions - product and geography.
Product and geography are what make cream skimming work.
1neutral
0
you know during the season and i guess at at your level uh you lose them to the next level if if they decide to recall the the parent team the Braves decide to call to recall a guy from triple A then a double A guy goes up to replace him and a single A guy goes up to replace him
You lose the things to the following level if the people recall.
0entailment
1
One of our number will carry out your instructions minutely.
A member of my team will execute your orders with immense precision.
0entailment
2
How do you know? All this is their information again.
This information belongs to them.
0entailment
3
my walkman broke so i'm upset now i just have to turn the stereo up real loud
I'm upset that my walkman broke and now I have to turn the stereo up really loud.
0entailment
5
(Read for Slate 's take on Jackson's findings.)
Slate had an opinion on Jackson's findings.
0entailment
7
At the end of Rue des Francs-Bourgeois is what many consider to be the city's most handsome residential square, the Place des Vosges, with its stone and red brick facades.
Place des Vosges is constructed entirely of gray marble.
2contradiction
9
I burst through a set of cabin doors, and fell to the ground-
I burst through the doors and fell down.
0entailment
10
Fun for adults and children.
Fun for only children.
2contradiction
11
Issues in Data Synthesis.
Problems in data synthesis.
0entailment
15
well you see that on television also
You can see that on television, as well.
0entailment
16
Vrenna and I both fought him and he nearly took us.
Neither Vrenna nor myself have ever fought him.
2contradiction
17
This analysis pooled estimates from these two studies to develop a C-R function linking PM to chronic bronchitis.
The analysis proves that there is no link between PM and bronchitis.
2contradiction
18
He turned and smiled at Vrenna.
He smiled at Vrenna who was walking slowly behind him with her mother.
1neutral
19
We sought to identify practices that were commonly implemented by the agencies within the past 5 years.
We want to identify practices commonly used by agencies in the last 5 years
0entailment
20
The other men shuffled.
The other men were shuffled around.
0entailment
21
States must show reasonable progress in their state implementation plans toward the congressionally mandated goal of returning to natural conditions in national parks and wilderness areas.
Itis not necessary for there to be any improvement.
2contradiction
22
well it's been very interesting
It has been very intriguing.
0entailment
23
He started slowly back to the bunkhouse.
He returned slowly to the bunkhouse.
0entailment
24
yeah well you're a student right
Well you're a mechanics student right?
1neutral
27
it really is i heard something that their supposed to be starting a huge campaign in New York about um child abuse and stopping child abuse and it's supposed to be like it's starting there supposed to be like a big nationwide campaign and you know so hopefully that will take off and really do something i don't know there's just
It's unfortunate that nobody is organizing a child abuse campaign.
2contradiction
28
Postal Service were to reduce delivery frequency.
The postal service could deliver less frequently.
0entailment
29
And in another shift in the economy, it was found that lamb could be raised more cost-effectively on lowland farms in part because of the richer, more nutritious grazing land available there and as a result Lakeland farms became less profitable.
Another shift in the economy was found to be more nutritious.
0entailment
30
The rule contains information collection requirements which will allow EPA to determine that detergent additives which are effective in controlling deposits are used and that emission control goals are realized.
The rule has data collection requirements which aid the EPA to realize their emission control goals.
0entailment
31
Almost every hill has to the northeast there's a Moorish fort; no fewer than four ruined fortresses guard the harbour entrance; and two more, still in good repair the Atalaya and Galeras castles protect the sea-front arsenal, of vital importance to Spain's military.
There are no castles Atalaya and Galeras.
2contradiction
32
Felicia's Journey takes place behind the eyes of its central a young Irish girl, Felicia, who crosses the sea to England in a hopeful quest to find the father of her unborn child; and the fat, middle-aged catering manager, Hiditch, who takes a paternal interest in the lass when it becomes clear that her young man has caddishly given her the slip.
The woman did not care where the man was as long as it was far.
2contradiction
34
In the crypt are interred the remains of Voltaire and Rousseau, Hugo and Zola, assassinated Socialist leader Jean Jaurys, and Louis Braille, the inventor of the alphabet for the blind.
The remains of Voltaire and Rousseau, Hugo and Zola, assassinated Socialist leader Jean Jaurys, and Louis Braille are all interred in the crypt.
0entailment
35
there's a uh a couple called um oh i'm going to forgot his name now uh Dirkson
I can't remember their name
0entailment
36
Poirot, I exclaimed, with relief, and seizing him by both hands, I dragged him into the room.
Poirot was now back and I was sorry that he would take over what I now considered my own investigation.
2contradiction
37
The word aswan actually means trade or market in ancient Egyptian, signifying its most pre-eminent activity.
Aswan's meaning has not changed over time.
2contradiction
38
well what the market can bear and
THe market can't bear any of it.
2contradiction
39
And to show just how fast Japan's new rulers were catching on, two punitive expeditions were launched against Korea and China in the grand manner of 19th-century gunboat diplomacy.
Japan's new rulers were catching on quickly.
0entailment
40
When the trust fund begins running cash deficits in 2016, the government as a whole must come up with the cash to finance Social Security's cash deficit by reducing any projected non-Social Security surpluses, borrowing from the public, raising other taxes, or reducing other government spending.
The public would generally prefer to see the government reduce its spending in other areas to finance Social Security.
1neutral
41
These organizations invest the time and effort to understand their processes and how those processes contribute to or hamper mission accomplishment.
These organizations invest lots of time to understand how some processes can contribute to or haampe
0entailment
42
and it is nice talking to you all righty
I talk to you every day.
1neutral
43
You have access to the facts.
The facts are accessible to you.
0entailment
44
Build environment Engineering Manufacturing Production (all rate tooling) (1st set of production tooling)
It is the first set of production tooling for manufacturing.
0entailment
45
yeah so um also of course they they can they join the they can always join the military service they are considered citizens i believe
They can't join the military service
2contradiction
47
She smiled back.
She was so happy she couldn't stop smiling.
1neutral
48
According to the Office of the Actuary at the Health Care Financing Administration, the estimated net present value of future additional resources needed to fund HI benefits alone over the 75 years is $4.
The net present value of future additional resources for funding HI benefits was $4.
0entailment
49
How do we fix this?'
Can we fix this?
0entailment
51
but that takes too much planning
It doesn't take much planning.
2contradiction
52
I did not mention Monica in my lecture, but the first question I was asked was how President Clinton could do his job with all the distractions caused by the Monica Lewinsky affair.
They wanted to get through the lecture without any problems.
1neutral
53
Because it plays on my childhood imagination.
The art plays on my young imagination.
1neutral
55
it's slow it's uh there are many better machines on the market right now for
This is the fastest machine, you won't find a better machine.
2contradiction
56
Consider the United States Postal Service.
Forget the United States Postal Service.
2contradiction
58
In full costume.
He is wearing a mascot costume.
1neutral
59
Although your journey is going to be difficult and at times you will doubt yourselves, the benefits to clients are well worth it and the satisfaction of hearing even the most recalcitrant of adversaries say that the new system is better than the one that went before is deeply rewarding.
The new system appears far more complex, but ultimately easier and more thorough.
1neutral
60
Take a remarkable statistic that Shesol cites but lets pass relatively unexamined.
They had data that was very relevant but under used.
0entailment
61
Get individuals to invest their time and the funding will follow.
If individuals will invest their time, funding will come along, too.
0entailment
63
The man on the ground thinks for a moment and yells back, You must work in management.
There was no one on the ground, man or woman.
2contradiction
65
If, however, the evaluation question requires GAO to report on how satisfactory progress is or the reasons for problems in implementation, the more staff who can be on site over time, with the ricHe'st or thickest base for examining the situation as the many people involved see it, the sounder our causal conclusions and subsequent recommendations will be.
If the GAO has to report on the progress, the recommendations will be much poorer quality.
2contradiction
66
The economy could be still better.
The economy has never been better.
2contradiction
67
Hills and mountains are especially sanctified in the cult of Jainism.
The cult of Jainism hates nature.
2contradiction
68
The famous tenements (or lands ) began to be built.
The land remained deserted.
2contradiction
69
Entreaties to the apartment's owner have gone nowhere.
The apartment's owner is very responsive.
2contradiction
72
Mrs. Cavendish is in her mother-in-law's room.
Mrs. Cavendish has left the building.
2contradiction
73
But I thought you'd sworn off coffee.
I thought that you vowed to drink more coffee.
2contradiction
74
We stink all the time.
We always stink.
0entailment
75
In an increasingly interdependent world, many pressing problems that affect Americans can be addressed only through cooperation with other countries.
We should be independent and stay away from talking and working with other nations.
2contradiction
76
The rule requires broadcasters to maintain a file for public inspection containing a Children's Television Programming Report and to identify programs specifically designed to educate and inform children at the beginning of those programs and to furnish such information to the publishers of program guides.
The rule makes broadcasters keep a file about children's television programming.
0entailment
78
well because how how hot i mean like like in the coldest that it gets in winter down there how much is it
It's hot all the time where I live, including winter.
1neutral
79
Each of the men wore leather armor and dressed in the style of heavy riders.
The men were naked.
2contradiction
80
The man should have died instantly.
The man was perfectly fine.
2contradiction
81
Just as the Daily Worker and New Masses , socialist papers from the 1930s, were peppered with citations of Marx and Engels, Educational Liberator is peppered with references to their libertarian equivalents--Friedrich Hayek and Ludwig von Mises.
Citations from Marx were used in socialist papers in the 30s.
0entailment
83
No matter what I ended up doing, I think I would have found some way within that profession to do public work, because that's what I was taught, Zelon said.
This is like using the postition as a banker to help make regulations better for the public.
1neutral
86
Clark also expressed the hope that he and Redgrave could continue with their marriage.
Clark hoped that he could continue their marriage.
0entailment
87
The poverty, for instance, does not create the sense of shame as it does for people who live in Western countries.
Poverty doesn't create a sense of shame in any country.
2contradiction
88
they supplied some uh you know they were some groups uh the Vicksburg if you ever get a chance to go over to Vicksburg the battleground at Vicksburg uh there's an area there where there was uh some Texas uh groups and they had an interesting time there
There were also North Carolina troops at Vicksburg.
1neutral
90
which i mean i think it should be anyway
I don't think it should be that way
2contradiction
91
Also, compensation committees need to understand the implications of compensation to provide incentives for management to do the right thing for the company and its shareholders versus themselves.
The implications of compensation to provide incentives for management to do the right thing for the company and its shareholders versus themselves should always be ignored by compensation committees.
2contradiction
92
The fourth-century Roman emperor Gratianus was an early visitor, followed much later by such luminaries as the Romantic poet Alphonse de Lamartine, Queen Victoria, Saint-Sa?«ns, and Rachmaninov.
Gratianus was a Roman empoeror.
0entailment
93
So, I went to court by myself and told them the truth, but it didn't do me any good.
I went and told the truth at court but it didn't do me any good.
0entailment
94
Click More Links (on the right-hand side under Miscellaneous), and from
There are no links to click under Miscellaneous.
2contradiction
95
no oh no oh well take care
Bye for now.
0entailment
97
'Hello, Ben.'
I ignored Ben
2contradiction
98
how can you prove it
Can you tell me how to prove it?
0entailment
99
It displays all kinds of vehicles, from the coach that carried Napoleon to and from Moscow in 1812 to a splendid 4-horsepower Renault car from 1904 and other turn-of-the-century classics.
The exhibition only displays cars from the 2000s.
2contradiction
100
According to this executive, the town hall meetings helped improve his customer satisfaction levels because veterans identified the concerns that were most important to them, gained direct access to the VBA employees working on their benefit claims, and were better able to understand the claims process.
The own hall meetings helped improve his customer satisfaction levels, said the executive.
0entailment
101
well see that isn't too bad a couple hours
I don't like waiting, not even for 10 minutes.
2contradiction
102
In the east, Christian forces were in retreat as a new Moslem power began to rise.
Although the Christian forces were stronger, they were in retreat.
1neutral
103
We have been told on reasonably high authority that true charity vaunteth not itself
True charity vaunts itself, we have been told.
2contradiction
104
yeah and what to do when they're faced with it
They'll face peer pressure and will need to say no.
1neutral
105
The With attorneys one year out of graduate school facing an average debt of just less than $90,000 and starting salaries at legal aid organizations averaging $31,000, they couldn't afford the job.
New attorneys have massive law school debt.
0entailment
106
Here you'll see a shrunken head, a two-headed goat, and a statue of Marilyn Monroe made of shredded money, among other curiosities.
One of the curiosities is a two-headed goat.
0entailment
107
Thursday's judge, the Honorable Charles Adams of the Coconino County Superior Court, agreed, but highly discouraged self-representation.
Self-representation was encouraged by the Honorable Charles Adams.
2contradiction
108
He mostly hangs out with a group of older, Southern black men, who call him Jumper and Black Cat.
The group of guys he tends to hang out with gave him the nickname Jumper.
0entailment
109
This elegant spa town on the edge of the Lac du Bourget has offered cures for rheumatism and other ailments for centuries.
The town was only established in the past fifty years.
2contradiction
110
right right well it's it's a beautiful city and but the problem is like first example when i was young they they took me to Las Vegas and that was the most boring place on earth
I think Las Vegas is the most boring place I know.
0entailment
111
i'm going to be able to make my own clothes
I will be able to make my own clothing.
0entailment
113
The river plays a central role in all visits to Paris.
The river is central to all vacations to Paris.
0entailment
114
they're supposed to be bilingual
They're supposed to be bilingual
0entailment
115
Oh, Czarek, but will happen to us?
Czarek was not asked any questions on that day.
2contradiction
116
In any case--it's over."
It is over.
0entailment
119
The herbal extract shows success in treating mild dementia and preventing Alzheimer's memory loss.
The herbal extract is not very good at treating dementia.
2contradiction
120
After a lightning victory over the ill-prepared French armies, the Germans marched on Paris and laid siege to the city, which finally capitulated in January 1871 in the face of dwindling food supplies.
Paris was unhappy to have to surrender to the Germans.
1neutral
121
Critical to the product development process were activities that enabled the capture of specific design and manufacturing knowledge and decision reviews to determine if the knowledge captured would support the increased investment necessary to move to the next development phase or into production.
The knowledge captured would determine if the next step was further development or production.
0entailment
123
Said I was foolishly proud.
My pride was foolish not well placed.
0entailment
124
The word had filled his head as though the girl had whispered directly into both ears.
The girl spoke very quietly.
0entailment
125
and a lot of their rules and regulations aren't real clear so we have our manager of environmental who assist the TACB which is located in Austin in writing and hey look what we've done here at TI
There rules are not really clear so we ask our manager for help.
0entailment
126

Dataset Card for Bias-amplified Splits for MultiNLI

Dataset Summary

Bias-amplified splits is a novel evaluation framework to assess model robustness, by amplifying dataset biases in the training data and challenging models to generalize beyond them. This framework is defined by a bias-amplified training set and a hard, anti-biased test set, which we automatically extract from existing datasets using model-based methods.

Our experiments show that the identified anti-biased examples are naturally challenging for models, and moreover, models trained on bias-amplified data exhibit dramatic performance drops on anti-biased examples, which are not mitigated by common approaches to improve generalization.

Here we apply our framework to MultiNLI, a crowd-sourced collection of 433k sentence pairs annotated with textual entailment information.

Our evaluation framework can be applied to any existing dataset, even those considered obsolete, to test model robustness. We hope our work will guide the development of robust models that do not rely on superficial biases and correlations.

Evaluation Results (DeBERTa-large)

For splits based on minority examples:
Training Data \ Test Data Original test Anti-biased test
Original training split 91.1 74.3
Biased training split 88.7 57.5
For splits based on partial-input model:
Training Data \ Test Data Original test Anti-biased test
Original training split 91.1 81.4
Biased training split 89.5 71.8

Loading the Data

from datasets import load_dataset

# choose which bias detection method to use for the bias-amplified splits: either "minority_examples" or "partial_input"
dataset = load_dataset("bias-amplified-splits/mnli", "minority_examples")

# use the biased training split and anti-biased test split
train_dataset = dataset['train.biased']
eval_dataset = dataset['validation_matched.anti_biased']

Dataset Structure

Data Instances

Data instances are taken directly from MultiNLI (GLUE version), and re-split into biased and anti-biased subsets. Here is an example of an instance from the dataset:

{
  "idx": 0,
  "premise": "Your contribution helped make it possible for us to provide our students with a quality education.",
  "hypothesis": "Your contributions were of no help with our students' education.",
  "label": 2
}

Data Fields

  • idx: unique identifier for the example within its original data splits (e.g., validation matched)
  • premise: a piece of text
  • hypothesis: a piece of text that may be true, false, or whose truth conditions may not be knowable when compared to the premise
  • label: one of 0, 1 and 2 (entailment, neutral, and contradiction)

Data Splits

Bias-amplified splits require a method to detect biased and anti-biased examples in datasets. We release bias-amplified splits based created with each of these two methods:

  • Minority examples: A novel method we introduce that leverages representation learning and clustering for identifying anti-biased minority examples (Tu et al., 2020)—examples that defy common statistical patterns found in the rest of the dataset.
  • Partial-input baselines: A common method for identifying biased examples containing annotation artifacts in a dataset, which examines the performance of models that are restricted to using only part of the input. Such models, if successful, are bound to rely on unintended or spurious patterns in the dataset.

Using each of the two methods, we split each of the original train and test splits into biased and anti-biased subsets. See the paper for more details.

Minority Examples

Dataset Split Number of Instances in Split
Train - biased 309873
Train - anti-biased 82829
Validation matched - biased 7771
Validation matched - anti-biased 2044
Validation mismatched - biased 7797
Validation mismatched - anti-biased 2035

Partial-input Baselines

Dataset Split Number of Instances in Split
Train - biased 309873
Train - anti-biased 82829
Validation matched - biased 7745
Validation matched - anti-biased 2070
Validation mismatched - biased 7758
Validation mismatched - anti-biased 2074

Dataset Creation

Curation Rationale

NLP models often rely on superficial cues known as dataset biases to achieve impressive performance, and can fail on examples where these biases do not hold. To develop more robust, unbiased models, recent work aims to filter bisased examples from training sets. We argue that in order to encourage the development of robust models, we should in fact amplify biases in the training sets, while adopting the challenge set approach and making test sets anti-biased. To implement our approach, we introduce a simple framework that can be applied automatically to any existing dataset to use it for testing model robustness.

Annotations

Annotation process

No new annotations are required to create bias-amplified splits. Existing data instances are split into biased and anti-biased splits based on automatic model-based methods to detect such examples.

Considerations for Using the Data

Social Impact of Dataset

Bias-amplified splits were created to promote the development of robust NLP models that do not rely on superficial biases and correlations, and provide more challenging evaluation of existing systems.

Discussion of Biases

We propose to use bias-amplified splits to complement benchmarks with challenging evaluation settings that test model robustness, in addition to the dataset’s main training and test sets. As such, while existing dataset biases are amplified during training with bias-amplified splits, these splits are intended primarily for model evaluation, to expose the bias-exploiting behaviors of models and to identify more robsut models and effective robustness interventions.

Additional Information

Dataset Curators

Bias-amplified splits were introduced by Yuval Reif and Roy Schwartz from the Hebrew University of Jerusalem.

MultiNLI was developed by Adina Williams, Nikita Nangia and Samuel Bowman.

Citation Information

@misc{reif2023fighting,
    title = "Fighting Bias with Bias: Promoting Model Robustness by Amplifying Dataset Biases",
    author = "Yuval Reif and Roy Schwartz",
    month = may,
    year = "2023",
    url = "https://arxiv.org/pdf/2305.18917",
}

Source dataset:

@InProceedings{N18-1101,
  author = "Williams, Adina
            and Nangia, Nikita
            and Bowman, Samuel",
  title = "A Broad-Coverage Challenge Corpus for
           Sentence Understanding through Inference",
  booktitle = "Proceedings of the 2018 Conference of
               the North American Chapter of the
               Association for Computational Linguistics:
               Human Language Technologies, Volume 1 (Long
               Papers)",
  year = "2018",
  publisher = "Association for Computational Linguistics",
  pages = "1112--1122",
  location = "New Orleans, Louisiana",
  url = "http://aclweb.org/anthology/N18-1101"
}
Downloads last month
1,736
Edit dataset card